Mutual Information-Based Counterfactuals for Fair Graph Neural Networks
Long Fei Chen, Wenzhuo Chen, Jiayi Li, Jiahui Yu · 2024
Graph neural networks (GNNs) have become a promising paradigm in the field of node representation learning, owing to their capacity to capture both node features and their relationships. However, GNNs could suffer from algorithmic discrimination in the context of high-stakes applications. To facilitate the algorithmic fairness of GNNs, several studies have been proposed for fair GNNs. However, a primary issue of fair GNNs is that ensuring algorithmic fairness could impair the node classification performance. In order to better balance algorithmic fairness and node classification performance, this work presents a mutual information-based counterfactual generation algorithm for fair GNNs. Specifically, the counterfactuals of the graph data are generated by a fairness adversarial loss, thus guiding the GNNs towards fair ones. Meanwhile, a mutual information-based mechanism is designed to minimize the discrepancy between the representations from the original graph and the counterfactuals, thus maintaining the node classification performance. Extensive experimental studies have been performed on three graph datasets to verify the efficacy of the proposed method.